“The Bots Are Here. Now What?” How Knowledge Management Became the Key to Powering GenAI Solutions

1. Introduction: The Arrival of the Bots

As far as Machine Learning (ML) was concerned, the years around 2020 were spent sharpening the tools to continue to do the same type of work, but more professionally. LinkedIn was full of articles on “what is the difference between supervised and unsupervised learning” and colleagues were discussing whether Machine Learning Engineer and Data Scientist were the same role. At the time, much of the operations part of ML were shaping up, giving us the MLOps we all now know. Let’s just say that ML maturity is pretty high right now.

Fast forward, from 2023 to now (July 2025) things have changed. The future is here, and the robots have arrived. Online, I see a lot of focus on the quantifiable aspects (e.g. evals where model A is 5% better than model B, hence model A is king and model B is a jester). Initially, I tried to keep up with news and comparisons, but I have to admit that since Llama 3 was released I stopped paying attention to the LLM race. Sure, open, cheaper and faster is always welcome, but I can tell that the bottleneck has shifted from the quality of the models to us. The robots have arrived.. and we are not ready for them.

This realisation came when I started working on a chatbot project in 2024. I was lucky to have the opportunity to test out the freshly released AIBots (a chatbot-builder-platform, courtesy of GovTech’s Government Digital Products (GDP)) and was able to build a proof of concept in literally under 15 minutes. The team at AIBots did an outstanding job then and have been able to continuously top themselves ever since. As a user of AIBots, I felt that by abstracting away the hardening, model hosting, knowledge chunking, embedding and Retrieval Augmented Generation (RAG) systems, all I was left with was the keys to a spaceship. I only needed to provide the “fuel”, which was the knowledge, and AIBots would handle the rest. Surely the knowledge was the easy part. Right?

2. The Forgotten Discipline: Knowledge Management

TL;DR: Enterprise knowledge management is not a new problem, but thanks to GenAI, motivation to address it is now sky-high.

While refining my proof of concept (which took longer than 15 mins) I got greedy and the knowledge I tried to squeeze in started growing and so did my struggle to organise it. It was during the recycling of a mixture of documents in various formats, that I started to discern the unique qualities of the task itself. So, I did a little digging on the internet. Turns out Knowledge Management is a field of work (the term only coming out in the 90s), but the practice has been there since we could paint in caves.

And, both from what I was experiencing and from reading online, let me tell you this: knowledge management is hard too. People seem to have always thought of the idea of knowledge management as valuable but largely neglected the implementation. “The search doesn’t work well”, “the content is stale”, “there is no ownership on the documents”. These were all problems I was facing in my little proof of concept too!. How to scale up?

3. Why KM Matters More Than Ever

Tl;DR: Taking the coding part seriously and not the knowledge management seriously will simply not work for real world usage.

Like Cassie says, a chatbot can only harm you if you let it. But have I mentioned agents? It’s been proclaimed that 2025 is the year of agents. Agents who can take actions and are earnest for detailed SOPs, corner-case catching instructions and a wide range of examples of “what good looks like”. Yet, as I witnessed these powerful LLMs and agents figuratively marching towards our project, we found ourselves squat-stirring at a stew of stale Power Points, PDFs and various chunks of text that I copy pasted from my company’s intranet into a variety of TXT files. We were not ready.

And that’s the point I’m hoping to make here. Your solution leveraging a bot/agent/whateverAI is as good as the content that powers it. The LLMs we have now are more than sufficient for the task. The vector databases that power the similarity search are perfectly well suited to pull out the correcttop-K references, that is, if you have those references in the first place. You can only do useful RAG if you have something worth R-ing.

Realising that the LLMs were not going to do most of the heavy lifting, we changed gears dramatically. We decided that if our chatbot was going to work effectively, we needed to treat all parts of the process with equal importance. If Machine Learning could have an MLOps, we were going to have a KnowledgeOps. And aside from generic statements and frameworks, we couldn’t find a specific stack or SOP to implement online, so we made our own.

4. Our KnowledgeOps

TL;DR: You cannot wing it with knowledge management. Define an SOP and stick to it.

It goes without saying that what works for us might not work for you. However, these recommendations should be implementable in whatever tech stack you have at your workplace. Our process is presented in Figure 1 below.

• Markdown: Simple and plain. Transparent and less prone to malicious prompt injections. You can perform RAG based on Markdown chunks, which gives a visually-friendly retrieval flexibility that a fixed chunk-size never could.
• Version Control with Git: When we keep a development and production version of our markdown files, we might end up with multiple versions of docs and rounds of revision. How is this not like code? We could even do unit tests against the knowledge, if we liked.
• Development and Production Bots: Never test knowledge changes in production. Do I need to say more?
• Review Cycles and Hotfixes: You need to include metadata like “date_last_updated” or “content_status”. This will help you orchestrate periodical freshness updates and checks. Even if nothing has changed, your user will appreciate that fresh timestamp. There will be a need for hotfixes, so do plan for them too.
• Testing: Remember when you thought that explaining precision and recallwas complicated? Simpler times indeed. What works for us is coming up with a set of questions that cover all important domains in the knowledge base. As we add more knowledge, we add more questions and involve experts who can assess the veracity of the answers to those questions. After various scorings and quantifications, we find ourselves satisfied with knowing the bot said either “correct information” or “abstained from answering” most of the time. Your boss can decide what % “most of the time” should be, provided that the number is strictly less than the FTE (Full Time Equivalent a.k.a workload) % of the knowledge curator.
• Monitoring: Once your bot is out in the wild, you must observe its interaction with the humans it meets. Here’s where having access to anonymised interaction is golden. You get to observe how people engage with your bot organically. Do they expect the bot to know a domain you haven’t covered? Do they know how to prompt the bot? Has someone found your bot’s Achilles’ heel? AIBots platform offers this feature and without it, we wouldn’t know what’s going on.

5. The Role of the Knowledge Curator

TL;DR: If you have a data product that requires someone to make it work, you need to make someone responsible and empowered to own it.

Maybe you are reading this and thinking: “This is all fine and dandy, but who’s going to be doing all this work?”. Well, a knowledge manager. Sometimes known as a librarian, curator or information organiser. Someone needs to fuel the spaceship so the rest can enjoy the ride. Let’s talk a bit more about this someone. Here are some traits that they should possess:

• A Steward: This job role doesn’t have to be a full time position but needs someone who is explicitly assigned the responsibilities and is well empowered to carry them out. If your subject matter expert won’t collaborate with the knowledge manager, the role is not set for success. And that goes for your AI solution too.
• Responsibilities: Define and maintain a style, strive for clarity and simplicity, deduplicate relentlessly, maintain freshness and metadata hygiene. You will also have to plan for new content intake schedules (which involves testing, development and deployment). Housekeeping too.
• Skillset: Software development background is certainly not a requirement, but it can help get the intuition and tools quickly. More importanter, is excellent written communication, social predisposition to pick the brains of subject matter experts and vanilla project management.

And just like that, for all the AI Doomersayers, we’ve just created a new job role for a human thanks to GenAI. Didn’t see that one coming.

6. Conclusion: What should you do?

TL;DR: Big knowledge management requires a plan and collaboration. Do you have one?

If everything your bot has to say can fit in a couple of A4 pages, you probably don’t need any of this. But, if you are covering multiple complex topics, that require various degrees of depth and abstractions, and the span of those topics is growing, you will need to invest some effort in managing your knowledge.

An important byproduct of collaborative knowledge management, is that the process of organising the knowledge helps teams better understand the problem. There’s a company that approaches this by having experts observe the challenges their users face and has them hammer out blog posts that can be used as knowledge for chatbots. By releasing and editing existing blogs in response to user’s challenges or ahead of new launches, they set their bot for success, and in this way their company too.

Your knowledge, just like the rest of your data, is an important resource and one that can make or break your GenAI solution. If you are serious about LLMs, Context Engineering and Agents, you need to be serious about knowledge management. LLMs may have read every article on Wikipedia and Reddit, but the realm of your solution is specific and if you don’t build a fence around your context in crystal clear fashion, you will get generic slop. It’s not a bug or a hallucination, it’s how it works. The more you treat your LLM as the smart parrot that it is and less like a knowledgeable reputable source, the more you will see the benefits of having a knowledge base to RAG-on.

The best time to set up your knowledge base was 5 years ago, the next best time is now. Define your scope, empower your knowledge manager and get to work on your knowledge management today.

7. MY templates are YOUR templates

Here’s a skeleton of what our documents look like. Again, your version might require adjustments but keep these content bits below as suggestions:

- -
title: "[Document File Name]"
date_last_updated: "YYYY-MM-DD"
content_status: "[DRAFT/APPROVED/ARCHIVED]"
- -

# Executive Summary
<! - AI-generated executive summary of the document content →


# Main Content
<! - Brief overview of the sections of the document and how they connect to one another. It's ok to be redundant, remember that the semantic search will perform better if you name your entities consistently. →

## Section 1
<! - Content here →
<! - URLs using markdown like [THIS](link) →

### Subsection 1.1
<! - Content here →
## Section 2
<! - Content here →ma

8. Epilogue: Useful tips

Here are the two most important tips:

1. Public sector work requires very high accuracy. The advice and guidelines on prompt engineering, user query and knowledge management all help with reducing LLM hallucinations. But don’t forget the most important thing you can do: managing expectations. No one should expect perfect answers from an LLM.
2. Remember that nobody wants to converse with a chatbot if they can avoid it. Especially at work. Make sure people get what they need as painlessly as possible.

Here’s a non-ranked list of tips and learnings we found useful. Hope you can find something that helps you too:

• If your solution is intra-company facing and you have an intranet, work to make your intranet the knowledge base. Nothing beats providing links to “the source” after every statement of the bot.
• Avoid non-text references. If you have a very cool chart, you will have to describe it and provide a link to it. Large Multimodal Models (LMMs) are not as reliable, lightweight and easy to handle as LLMs. At least for now.
• Like with any product, define your user and build a fence around its needs. This will in turn help narrow down the scope of the knowledge that you need to manage.
• Don’t try to automate something that a human couldn’t do (with enough time and motivation). If you are asking your chatbot questions that would make a subject matter expert pause and hesitate, your bot will fare far worse.
• A lot of people are still learning how to prompt adequately. You can use some of your benchmark questions as “suggested questions” for them to get started. These also help illustrate the boundaries of the knowledge base.
• Make your knowledge base available to your users. We position our chatbot as a guide, whereas the knowledge base itself is the authoritative source of information. Providing clickable reference links to the knowledge after a bot finishes the statement really helps solve a lot of questions.
• Abstaining from answering is not a failure. Letting the bot go ahead and hallucinate is.
• Monitor chat sessions. This will give you an indication of topics that could use improvement or new knowledge bits to include in the future.
• Use the long context that your LLM provides for key information (e.g. customer service phone number). This is information that should accompany every query and you definitely don’t want to get wrong. Popular company acronyms should go here too.
• Deduplicate aggressively. Look for key terms and tools and make sure they are introduced exactly once. Others should reference them.
• Use powerful markdown tools that offer backlinks and documentation linking (e.g. Obsidian) to monitor duplication and coverage of your knowledge base.
• For each document, use an LLM to create a brief executive summary. If the summary doesn’t capture the knowledge correctly, chances are that the RAG will fail to interpret correctly too, so you’ll need to revise the clarity of your doc. Include the final executive summary at the top of each document.